LLM hallucination depends on ambiguity of the prompt
The more I explore LLMs, the more I feel that hallucination is deeply connected to ambiguity.
People usually think hallucination only happens when the model invents fake facts.
But even normal language can create uncertainty.
Example:
“The cat is sitting on the soft mat and it is soft.”
The word “it” itself is ambiguous.
And now the model has to infer meaning from probability, context, and prior patterns.
What’s interesting is that humans also communicate this way constantly. Language is compressed and incomplete by default.
The difference is that humans are grounded in reality through experience, while LLMs are grounded mostly in language patterns.
Which is probably why ambiguity becomes such a big issue in long reasoning chains and complex prompts.